measuring crosswalk safety at non ...n.saunier.free.fr/saunier/stock/fu16crosswalk.pdf11 midblock...

1

MEASURING CROSSWALK SAFETY AT NON-SIGNALIZED CROSSINGS DURING

NIGHTTIME BASED ON SURROGATE MEASURES OF SAFETY: A CASE STUDY IN

MONTREAL

Ting Fu, Corresponding Author, PhD Candidate

Department of Civil Engineering and Applied Mechanics, McGill University

Room 391, Macdonald Engineering Building, 817 Sherbrooke Street West

Phone: (514) 632-1633

Montréal, Québec, Canada H3A 0C3

Email: [email protected]

Luis Miranda-Moreno, Associate Professor

Department of Civil Engineering and Applied Mechanics, McGill University

Room 268, Macdonald Engineering Building, 817 Sherbrooke Street West

Montréal, Québec, Canada H3A 0C3

Phone: (514) 398-6589

Fax: (514) 398-7361


Nicolas Saunier, Associate Professor Department of Civil, Geological and Mining Engineering

Polytechnique Montréal, C.P. 6079, succ. Centre-Ville

Montréal, Québec, Canada H3C 3A7

Phone: (514) 340-4711 x. 4962


Word count: 4250 words + 11 tables/figures x 250 words (each) + references (500 words) = 7500 words

November 15th, 2015

mailto:[email protected]



Fu, Miranda-Moreno, Saunier 2

ABSTRACT 1

This paper proposes a methodology to evaluate crosswalk safety at nighttime using surrogate 2

safety measures and thermal video sensors. The methodology is illustrated using two non-3

signalized crosswalk locations in downtown Montreal, Quebec. Video data recordings from the 4

thermal camera were used to compare nighttime and daytime safety conditions using different 5

surrogate safety measures including vehicle crossing speed, post encroachment time (PET), as 6

well as the yielding compliance and the conflict rate. A new way of measuring pedestrian 7

exposure is also proposed which excludes non-interacting road users. A thermal camera was used 8

in an effort to alleviate issues pertaining to low visibility at night for video analysis when road 9

users, especially pedestrians, are difficult to detect and track. The results showed that the 10

proposed thermal-video-based surrogate safety methodology is effective to collect and analyse 11

pedestrian-vehicle interactions at night regardless of lighting conditions. From the study 12

crossings, results also showed that that average vehicle crossing speeds and percentage of 13

dangerous conflicts were higher during nighttime compared to daytime, indicating that 14

pedestrians were at higher risks during nighttime. The proposed methodology can be used to 15

evaluate the performance of different crosswalk treatments on pedestrian safety at night. 16

17

Keywords: Nighttime Safety, Crosswalk Safety, Video Analysis, Thermal Videos, Surrogate 18

Safety Measures, Interactions, Risk Rate, Post Encroachment Time 19

20


INTRODUCTION 1

Pedestrian safety has become a priority for many cities due to increased awareness of their 2

vulnerability compared to other road users. In the U.S, 14 % of total road crash fatalities were 3

pedestrians in 2013 (1). Meanwhile 15.6 % of road crash fatalities in Canada were pedestrians in 4

2013 (2). Studies indicated that nearly half (46 %) of pedestrian fatalities in the US (3) and 57 % 5

of pedestrian fatalities in Ontario, Canada (4) occurred at nighttime. Most crashes happen when 6

pedestrians are crossing the streets when they are exposed to motorized traffic. A study in 7

Europe showed that roughly 31 % of all pedestrian victims of road crashes were injured on 8

marked crosswalks (5). Pedestrians are also vulnerable at locations with non-signalized 9

crossings. For instance, Hunter et al. found that 40 % of intersection crashes and 93 % of 10

midblock crashes occurred at non-signalized locations (6). Compared to daytime, there is less 11

motorized and pedestrian traffic at nighttime, which generally leads to higher vehicle speeds as 12

well as lower levels of driver awareness and attention. This difference in traffic and driving 13

behavior leads to an increase in crash frequency and severity especially when pedestrians are 14

involved (7) (8). In addition, Huang et al. point out that at nighttime, crosswalks and pedestrians 15

can be less visible for drivers to see in time for a stop (9). Crosswalk safety has been looked into 16

by numerous studies, and different treatments have been implemented and evaluated for different 17

crosswalk locations; however, evaluating the safety of the treatment is challenging, in particular 18

at night time, because of the sparse nature of the crash data and the lack of exposure measures 19

(e.g., count data during night time). Most often, short-term counts for safety analysis are taken 20

only during day time (10). 21

Moreover, the pedestrian safety literature has been built mainly through the use of 22

historical crash data, focusing on crash frequency and severity as direct measures for road safety 23

(11) (12) (13). However, vehicle-pedestrian crash data is not always available in sufficiently 24

large quantity and suffers from known problems such as low-mean small sample, underreporting, 25

mislocation and misclassification. Tarko et al. list the limitations of using crash data for road 26

safety analysis (14). In addition, the low mean problem (sparse nature of the crash data) can 27

represent a statistical issue when working with pedestrian crash data during nighttime, in which 28

given the low level of exposure, the mean number of crashes is typically very low. Using 29

historical data for pedestrian safety analysis requires long periods of observation (many years); 30

thereby recent treatments cannot be quickly evaluated from crash data due to the lack of after-31

treatment crash data (15). In order to overcome this problem, proactive methods have been 32

proposed that do not require waiting for crashes to happen. They rely on surrogate measures of 33

safety that may provide better and more precise alternative road safety indicators. 34

Surrogate safety measures that rely on automated video analytics are gaining increasing 35

popularity in road safety analysis for their various benefits (14) (15). Some studies have used 36

such measures for identifying risk factors or evaluating treatment effectiveness using a before-37

after or control-case study approach (15) (16) (17) (18) (19). St-Aubin et al. developed a 38

trajectory-based algorithm to measure Time-to-Collision (TTC) and carried out a Montreal case 39

study to evaluate a safety treatment on highway ramps (15). This work has been extended by 40

using realistic motion prediction methods (motion patterns learnt from observation) for the 41

evaluation of the safety performance of roundabouts (16). Despite the important developments 42

on surrogate safety analysis, there has been little nighttime safety evaluation using surrogate 43

measures. Among the reasons, one can mention the technological limitations of regular video 44

cameras (in the visible spectrum) that are unable to provide high quality data at night. 45

The objective of this work is to propose a surrogate safety methodology to quantify 46


pedestrian safety on crosswalks during nighttime using thermal video sensors. To get effective 1

video data under nighttime conditions, this paper used a thermal-camera based system; the 2

details of this system are reported in (20). The trajectories were then extracted from video data 3

and analyzed by calculating speeds of crossing vehicles and Post-Encroachment Time (PET). 4

This paper begins with a literature review; it then describes the thermal camera system 5

used to obtain usable nighttime data and measures for crosswalk safety based on video data. A 6

case study conducted as an example of using thermal videos for crosswalk safety evaluation at 7

night is presented. Finally, conclusions and future works are discussed. 8

9

LITERATURE REVIEW 10

Using traffic trajectory data obtained from video recordings is the most widely adopted method 11

for automatically calculating surrogate safety measures. Different trackers have been developed 12

and used to obtain trajectory data (21) (22) (23) (24) (25). Saunier et al. adapted this method to 13

intersections to track all road users by continuously detecting new features and adding them to 14

current feature groups (22) (23). An improved multiple object tracking system, named Urban 15

Tracker, was developed for tracking different types of road users in urban mixed traffic (25). In 16

addition, in order to count different road users in mixed traffic conditions and to identify 17

interactions based on their trajectories and between different types of road users, Zangenehpour 18

et al. developed a classification algorithm to distinguish between three types of road users: 19

pedestrians, vehicles and cyclists (26). The proposed classifier in this study uses the occurrence 20

area, speed distribution and presence (appearance) of the road users to classify them. The data 21

can then be used for surrogate safety analysis of the interactions between different road users. 22

The overall accuracy of the classification algorithm at intersections with high volumes and 23

mixed road user traffic was approximately 93 %. This algorithm was trained for thermal video in 24

(20) and results for mixed traffic conditions demonstrated an overall accuracy of 70 %. 25

Different studies have also used trajectory data for obtaining traffic information such as 26

volume, speed and conflict measures, which are fundamental for surrogate safety measures (15) 27

(16) (17) (26) (27) (28) (29). Laureshyn looked at different indicators in behavioral and road 28

safety research in terms of validity and reliability (30), and the indicators include time to 29

collision (TTC), post-encroachment time (PET), gap time (GP), encroachment time (ET), time 30

headway/time gap, compliance with the yielding rules and stop sign requirements and etc. 31

Different studies used different measures for different conditions. In (15) (16), St-Aubin et al. 32

computed TTC using the equations presented in (30) for highway safety. Tang and Nakamura 33

relied on PET for evaluating conflict severity at signalized intersections(31). For pedestrian 34

safety at crosswalks, PET has been widely used (17) (29). For instance, Alhajyaseen et al. (29) 35

used PET and vehicle speed at a crosswalk as validation parameters to assess pedestrian safety at 36

intersections. 37

Another important concept is the exposure of pedestrian to the risk of collision with 38

motor vehicles (32). Exposure is traditionally measured through the pedestrian and vehicle 39

volumes passing the area of interest, i.e. crosswalks for our study, or their product. But exposure 40

is a general concept that represents the opportunities or necessary conditions for a collision to 41

occur: it can be measured in various ways which depend on the purpose of the study. Pedestrians’ 42

exposure was already used in 1989 in a study of pedestrian safety at traffic signals using a 43

manual traffic conflict technique (TCT) (33). In 1998, Silcock et al (34) proposed a method that 44

used video recording as a data collection method to automatically extract data from video tapes 45

describing the number of crossing movements and pedestrian-vehicle interactions. However, the 46


definition of the conflicts (e.g. the threshold used on the surrogate measure of safety to 1

distinguish from other events) was not clarified (35). Exposure is generally used to calculate 2

pedestrian risks of collision with vehicles through crash or conflict ratios. The ratio is calculated 3

based on the number of crashes or conflicts over the total exposure, which reflects the probability 4

dimension of risk, i.e. the probability of a crash or conflict per unit of the chosen exposure. The 5

most recent work using surrogate safety analysis with rate calculations can be found in (26). In 6

this paper, the authors used the ratio of the total number of conflicts and severe conflicts divided 7

by the product of the pedestrian and vehicle volumes. Other indicators such as speed and 8

yielding compliance to evaluated crosswalk safety have been used extensively in similar studies. 9

Different measures of pedestrian exposure have been proposed (35). In the literature, the 10

number of pedestrian crossings (per hour), vehicle volume, or their product have been used as 11

the key indicators; however, these measures do not correspond to events where a pedestrian and 12

a vehicle may actually interact, i.e. they are close enough to each other at the site of interest that 13

they are at least aware of each other. There is a huge gap between the product of traffic volumes 14

and an actual interaction between a pedestrian and a vehicle. This gap is even larger during 15

nighttime in which pedestrian and vehicle flows are much lower and can present more temporal 16

variability. Vehicle-pedestrian interactions change from site to site and from time to time due to 17

many conditions at different sites. Besides, upstream signalization has a large impact on the 18

arriving time of the pedestrians and the vehicles, which also influences pedestrian exposure. All 19

these uncertainties may explain the low or unreported model fitness in past studies. All these 20

require proposing and testing existing and new exposure measures. 21

22

METHODOLOGY 23

The methodology consists of three key steps: thermal video data collection, trajectory extraction, 24

and computation of surrogate safety measures. 25

26

Thermal Video System, Object Tracking and Validation of Detection Performance 27

A thermal camera system was used for data collection. For details about the system and its 28

performance in nighttime conditions, one can refer to (20). The system components are presented 29

in FIGURE1 a). For field measuring purposes, the camera was mounted on an adjustable mast 30

against existing poles (i.e. lamppost or telephone pole) with an ideal coverage area and camera 31

angle. FIGURE 1 b) shows a sample snapshot from the thermal video which was taken at 32

nighttime where regular cameras in the visible spectrum fail to provide enough details about road 33

users because of the darkness, reflection, and shadow and glare from different light sources. 34

FIGURE 1 c) presents the issues of using regular videos for video data collection at night, and 35

how thermal video is not affected by these lighting issues. 36

37


1 a) Camera system and installation b) Sample of thermal video 2 3

4

5 c) Issues for using regular videos at night paired with the corresponding thermal videos 6 7

FIGURE 1 Thermal camera system and a comparison with regular videos at nighttime 8

9

Once video was collected, video data processing was carried out using the tracker in the 10

open source Traffic Intelligence project (22); as an outcome, road user trajectories were obtained. 11

The techniques used in the tracker are explained by Shi and Tomasi (21) and Saunier(23). (20) 12

has validated the performance of video analysis using thermal video for traffic data collection in 13

multi-modal environments in various lighting and temperature conditions, and has shown the 14

reliability of this technique. Compared with mixed traffic conditions at intersections, non-15

signalized crosswalks are much simpler because road users travel in fixed directions along fixed 16

segregated paths. Therefore, the performance of the tracker for detecting road users at crosswalks 17

is expected to be higher. This study uses the performance measures introduced in (20). Miss rate 18

was defined in (20) as “the proportion of road users whose movement is not captured by any 19

trajectories”, and “was used to quantify detection performance. For pedestrians, the detection 20

performance was evaluated at the group level, i.e. a group of pedestrians not tracked is counted 21

as one miss”. Precision and recall for detection are also reported. 22

23

amera

Issue 1: Headlight

reflection Issue 2:

Low visibility Issue 3:

Glare from light sources

Issue 4:

Shadow


Safety Measures 1

For evaluating the safety status of a crosswalk during night time, the following three measures 2

were defined. 3

4

Pedestrian-Vehicle Interaction 5

FIGURE 2 describes an interaction between a pedestrian and a vehicle at a crosswalk. PET is 6

defined as “the time gap” between two road users “arriving” and “leaving” the crossing area; 7

PET is used in this study as the surrogate safety measure for interactions between pedestrians 8

crossing the street and vehicles since their trajectories will always intersect and PET can thus 9

always be computed. The fact that PET may not be computed for some interactions is otherwise 10

a known shortcoming of that measure. Based on the road user classification and the trajectory 11

data of each road user, the PETs of pedestrian-vehicle interactions is calculated as: 12

13

𝑃𝐸𝑇 = {𝑇𝐶𝐹 − 𝑇𝑃, 𝑖𝑓 𝑇𝑃 < 𝑇𝐶𝐹

𝑇𝑃 − 𝑇𝐶𝑅, 𝑖𝑓 𝑇𝐶𝑅 < 𝑇𝑃

(1) 14

15

Where the notations are defined and illustrated in FIGURE 2: 𝑇𝑃 < 𝑇𝐶𝐹 indicates the situation 16

where the pedestrian arrives at the crossing area before the vehicle, while 𝑇𝐶𝑅 < 𝑇𝑃 means the 17

opposite. The study used the trajectory data to measure the PET between each pedestrian 18

crossing the street and vehicle crossing the crosswalk at a same time. This study used a computer 19

vision safety analysis tool to automatically calculate the PET values for each pair of interacting 20

vehicle and pedestrian. Interactions with PETs less than 5 seconds were considered as conflicts, 21

and those with PETs less than 1.5 seconds were defined as dangerous conflicts. For details about 22

the PET thresholds for the conflicts, see (19). With the computer vision software, pedestrians 23

may be tracked in a group, in which case only one interaction with the whole group will be 24

counted - in real situation, a small group pedestrians walking together could be regarded as one 25

road user as they have the same chance in interactions with passing vehicles. 26

27

28 29

FIGURE 2 Description of pedestrian-vehicle interactions at a crosswalk 30

31

Pedestrian Exposure Measure 32

Exposure measures, in most of the literature, are based on traffic volumes. Different exposure 33

TC0: time at when the vehicle is in the initial location

TCF: time at which the front of the vehicle reaches the crossing area

TCR: time at which the rear of the vehicle passes the crossing area

TP0: time when the pedestrian is in the initial location

TP: time at which the pedestrian reaches the crossing area

Tp

Crossing Area:

Intersection of Trajectories

TC0

TP0

TCR TCF

Vehicle Initial Location

Vehicle Location – Front of the

Vehicle Reaches the Crossing

Area

Vehicle Location – Rear of the

Vehicle Reaches the Crossing

Area


measures can be used depending on the purpose and can be considered in the traditional safety 1

hierarchy framework of surrogate safety analysis based on earlier work by Hydén among others 2

(36) illustrated in FIGURE 3, with collisions as the most severe events at the top and 3

undisturbed passages at the bottom. Using microscopic trajectory data, this work can measure 4

exposure at the level of road user interactions, when two road users are close enough in time and 5

space. This paper sets an arbitrary threshold of 20 s on PET for interactions considered as 6

exposure to pedestrian-vehicle collisions. 7

8

9 10

FIGURE 3 Pedestrian-vehicle interactions in the safety hierarchy (36) 11

12

Safety Measures 13

Safety measures are analyzed by visualizing the cumulative distribution functions (CDFs) and 14

the interaction rate based on the exposure measure defined above: this paper calculates the 15

interaction rate at crosswalks as the number of conflicts over the number of interactions used as 16

exposure. 17

18

Cumulative Distribution Functions (CDFs) 19

Visualization has been proved to be powerful for comparison purposes. CDFs are used in this 20

study as a way to understand pedestrian risk at crosswalk. FIGURE 4 demonstrates the principle 21

of this analysis method. The elevated line indicates a higher proportion of dangerous conflicts. 22

The grey line in the figure represents the dangerous conflict threshold and the right border of the 23

figure is the conflict threshold. This method of showing safety is intuitive; however, as (37) 24

illustrated, using cumulative distribution is not always conclusive. 25

26

Collisions

Light traffic conflicts

Undisturbed passages

Severe traffic conflicts

Potential traffic conflicts All interactions

(total exposure)


1 2

FIGURE 4 Illustration of different cumulative distribution functions 3

4

Conflict Ratio 5

Two conflict rates are used. For a given site i, the conflict rate (𝑅𝐶𝑖) is defined as the number of 6

pedestrian-vehicle conflicts, which are the interactions with PETs less than 5 s, divided by the 7

number of interactions with PET less than 20 s denoted 𝑁𝐸𝑖 (exposure). The dangerous conflict 8

rate (𝑅𝐷𝐶𝑖) is defined as the number of dangerous conflicts, which are the interactions with PET 9

less than 1.5 s, divided by the same exposure measure 𝑁𝐸𝑖 . 10

11

Other Safety Measures: Crossing Speed and Yielding Compliance 12

Crossing Speed 13

The crossing speed of vehicles passing the crosswalk was used as a safety measure in this paper. 14

Crossing speeds were automatically extracted from the videos through the computer vision 15

software and have been shown to be reliable in (20). A script was used for extracting velocities 16

and calculating the speeds for vehicles passing the crosswalk. FIGURE 5 presents how the 17

crossing speeds were extracted. A mask was prepared for the detection zone – the crosswalk in 18

this case, shown in FIGURE 5 a). In video collected from site 𝑖, for a certain vehicle 𝑗, 𝑗 =19

(1, … ,𝑁), where 𝑁 is the total number of vehicles, if its trajectory falls in the detecting zone in 20

video frame 𝑚, 𝑚ϵ(𝑝, 𝑝 + 1,… , 𝑞), its velocity 𝑣𝑖𝑗𝑚⃑⃑ ⃑⃑ ⃑⃑ ⃑⃑ is extracted and the instantaneous speed 21

𝑠𝑖𝑗𝑚 is calculated. The crossing speed is calculated by averaging the instantaneous speeds in 22

these frames, as presented in the following equation: 23

24

𝑠𝑖𝑗 =1

𝑞−𝑝+1∑ (𝑠𝑖𝑗𝑚)

𝑞𝑚=𝑝 (2) 25

26

The average crossing speed for site 𝑖 would be: 27

28

𝑠𝑖 =1

𝑁∑ (𝑠𝑖𝑗)

𝑁𝑗=1 (3) 29

30

Speed distributions and average crossing speeds of all the passing vehicles were compared. 31

32

Cu

mu

lati

ve

per

cen

tag

e (%

)

PET value

Dangerous conflicts

Conflicts

Dangerous conflict

threshold


1 a) Mask for detecting zone b) Sample of a trajectory 2

3

FIGURE 5 Sample of speed extraction through the computer vision software 4

5

Yielding Compliance 6

The law requires vehicles to yield to a pedestrian when he is starting or indicating the intension 7

to cross the road. In this case, yielding compliance refers to the rate of drivers’ yielding behavior 8

among the pedestrian-vehicle interactions which require the drivers to slow down or stop to give 9

pedestrians the right-of-way. Yielding compliance rate (𝑌𝐶𝑅 ) was calculated by manually 10

counting the vehicle yielding maneuvers. For site 𝑖, if a pedestrian arrives at the crosswalk before 11

a certain vehicle 𝑗 , the yielding behavior of this vehicle involved in an interaction with a 12

pedestrian can be quantified by the following measures 13

14

𝑋𝑖𝑗 = {1 𝑖𝑓 𝑡ℎ𝑒 𝑣𝑒ℎ𝑖𝑐𝑙𝑒 𝑦𝑖𝑒𝑙𝑑𝑠 𝑎𝑛𝑑 𝑔𝑖𝑣𝑒𝑠 𝑟𝑖𝑔ℎ𝑡 𝑜𝑓 𝑤𝑎𝑦 𝑡𝑜 𝑡ℎ𝑒 𝑝𝑒𝑑𝑒𝑠𝑡𝑟𝑖𝑎𝑛0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

, 𝑗 = (1,… ,𝑀𝑖) (4) 15

16

𝑌𝑖 = ∑ 𝑋𝑖𝑗𝑀1 (5) 17

18

𝑌𝐶𝑅𝑖 =𝑌𝑖

𝑀𝑖 (6) 19

20

Where 𝑀𝑖 is the total number of interactions between the crossing vehicle and the pedestrian 21

already starting or indicating his intention to cross and, to avoid a collision, at least one involved 22

road user must yield. 𝑌𝑖 is the total number of yielding drivers and 𝑌𝐶𝑅𝑖 is the yielding 23

compliance rate. 24

25

Validation of the Classification Tool for Pedestrian-Vehicle Interactions at Crosswalks 26

In order to calculate the PETs, a classification method is required to identify the vehicle-27

pedestrian interactions. A modification of a previously developed method for object 28

classification in video (26) was used in this study. The modification was done by changing the 29

image database for detecting road user presence in thermal videos (20). One can refer to (26) and 30

(20) for details. FIGURE 6 presents a sample snapshot of tracking and classification results. 31

In (20), the overall accuracy of the classification algorithm in terms of classification 32

performance measures has been shown to be over 80 % for mixed traffic with the average 33

precision of 70.9 % and the average recall of 99.5 % for vehicles, the average precision of 34

73.2 % and the average recall of 89.2 % for cyclists and the average precision of 98.6 % and the 35

Detecting Zone

– the Crosswalk

Position at Frame (p+1)

Position at Frame (p)

Position at Frame (p-1)

Position at Frame (q)

Position at Frame (q+1) Detecting

Zone


average recall of 72.0 % for pedestrians. While the rates are relatively high, they are not high 1

enough to conduct a safety analysis. However, with the simpler traffic conditions at non-2

signalized crosswalks, the performance of classification algorithm is expected to be better. 3

Similarly to (20), the classification performance was validated in terms of precision, recall and 4

overall accuracy, and was measured by extracting frames at every 10 consecutive seconds of 5

video. 6

7

8 9

FIGURE 6 Tracking & Classification process with video - Sample of tracking and 10

classification. The red line represents the trajectories of the moving objects up to the time 11

of the image; P stands for pedestrians and C stands for cars 12

13

CASE STUDY 14

15

Site Selection and Data Description 16

FIGURE 7 shows the locations of the selected sites. For testing the thermal camera system and 17

investigating the crosswalk safety, two crosswalk locations with different traffic and 18

environmental conditions were selected in downtown Montreal: 19

Site du Fort: This crosswalk is located on Rue du Fort at the intersection of Rue du Fort 20

and Rue Baile. It is a painted crosswalk crossing two one-way lanes and a median 21

between the two lanes. Since the left lane was observed to have very little traffic, only the 22

right lane was analyzed. Located on a secondary road, this site has a relatively low traffic 23

volume. 24

Site St-Laurent: The crosswalk is located on one of the main arteries in downtown 25

Montreal, Boulevard St-Laurent, at the intersection of Boulevard St-Laurent and Rue 26

Bagg. It is a painted crosswalk crossing two one-way lanes. This location is busier than 27

the du Fort site in terms of vehicular and pedestrian traffic. 28

29


1 2

FIGURE 7 Locations of the selected sites 3

4

For each site, thermal video data were collected in both daytime and nighttime 5

conditions. A total number of 16.8 hours of video data were collected. For comparison purposes, 6

all video data were collected in the same season with similar traffic, weather and road surface 7

conditions (i.e. collected on good weather conditions with bare pavement in winter). All the 8

videos were recorded during the afternoon peak period and at nighttime on weekdays when 9

higher crash rates were observed. Details of the video data are presented in Table 1. 10

11

Site: St-Laurent

Site: du Fort

crosswalk location


TABLE 1 Description of the Video Recorded from Each Site 1

Site name du Fort St-Laurent

Camera View

Total hour collected at night 4.6 hrs 6.0 hrs

Total hour collected during the day 2.6 hrs 3.6 hrs

Site name Date Period of a Day Time Duration (hour)

du Fort

11-Nov-2014 Day 03:00-04:40pm 1.6

Night 07:00-08:15pm 1.2

14-Jan-2015 Night 06:30-08:20pm 1.8

15-Jan-2015 Day 03:30-04:30pm 1.0

Night 07:00-08:40pm 1.6

St-Laurent

31-Dec-2014 Day 03:00-04:10pm 1.2

Night 06:30-09:15pm 2.8

01-Jan-2015 Day 02:00-04:30pm 2.4

Night 06:00-09:15pm 3.3

3

Detection and Classification Validation 4

The tracker and the classification algorithm were validated using 30 minute video samples from 5

each site. Results of the detection and classification performance are provided in Table 2. 6

7

TABLE 2 Classification Accuracy Validation Results 8

No. of

Presence

No. of Missed

/Miss Rate

Detection Performance Classification Performance

Precision Recall Precision Recall

du Fort (Video length of 30 minutes, from night, 14-Jan-2015)

Vehicle 68 1 / 1.7% 97.1% 97.1% 98.5% 98.5%

Pedestrian 60 1 / 1.6% 100% 81.1% 100% 98.3%

Overall 128 2 / 1.6% 91.4% 98.4% Global Accuracy = 97.7%

St-Laurent (Video length of 30 minutes, from night, 31-Dec-2014)

Vehicle 205 0 / 0% 97.6% 98.6% 94.8% 97.6%

Pedestrian 104 1 / 1.0% 100% 80.0% 100% 93.0%

Overall 309 1 / 0.3% 93.1% 98.4% Global Accuracy = 91.7%

9

Based on the results, the tracker and the classification algorithm worked almost perfectly 10

in detecting and classifying the pedestrians and vehicles at each crosswalk – very few misses and 11


around 95 % of precision, recall and global accuracy rates in most cases, except for the lower 1

recall values in detecting pedestrians (around 80 %) mainly resulting from the over-grouping of 2

pedestrians moving together (20). These values are much higher than those for mixed traffic 3

tested in (20), indicating the reliable performance of using the tracker and the classification 4

algorithm at crosswalks. The small portion of the misclassified road users could be easily 5

corrected in the output SQLite file. 6

7

Results and Analysis 8

The proposed methodology was applied to the selected sites and video data were processed. 9

Mean vehicle speed over the vehicle trajectory within the marked crosswalk area was calculated 10

as the crossing speed for each vehicle, and PETs between vehicles and pedestrians were 11

computed. For the du Fort site, an average of 319 vehicles and 161 pedestrians per hour were 12

detected during the 2.6 hours of video data collected in daytime conditions while a volume of 13

414 vehicles and 127 pedestrians per hour were detected from the 4.6 hours of video taken at 14

night. The St-Laurent site had a volume of 848 vehicles and 994 pedestrians per hour during 3.6 15

hours of daytime video recordings, and a vehicle flow of 833 vehicles and 407 pedestrians per 16

hour during 6.1 hours of nighttime video recordings. Table 3 presents a summary of the results of 17

different safety measures, which includes the vehicle crossing speed, vehicle yielding 18

compliance rate, exposure measured in the traditional way using the product of pedestrian and 19

vehicle volume, number of the conflicts, conflict rate, number of dangerous conflicts and rate of 20

dangerous conflicts, for both sites. FIGURE 8 presents the distributions of speeds and the CDFs 21

of PET for conflicts for both day and night. 22

Looking at FIGURE 8 a) and c) it can be observed that increases in the crossing speed 23

were detected at night for both sites. Also, from Table 3, for the crosswalk safety situation, the 24

average crossing speeds were found to be higher (by 9.3 % - 16.8 %) at the crosswalk of the du 25

Fort site at nighttime compared to daytime; for the St-Laurent site, the average crossing speeds 26

increase by around 30 % at night compared to daytime. Possible reasons for this observation 27

could be: 1) although the traffic flow is similar between afternoon peak hours and early nighttime 28

hours for both sites, the volumes at the second site are higher during the afternoon peak hours, 29

which leads to the congestion of the adjacent road segments; 2) during the afternoon peak hour, a 30

large number of vehicles are searching for parking spots and their parking maneuvers block the 31

traffic. This phenomenon is especially evident for the site of St-Laurent, where a pharmacy and 32

many restaurants are located. Many parking maneuvers were observed in the daytime while 33

fewer occurred at night; 3) Because of lower traffic volumes and less pedestrian activity at night, 34

drivers drive faster. This increase in the average crossing speeds of the passing vehicles at the 35

crosswalks at nighttime indicates that pedestrians are exposed to higher probabilities of severe 36

crashes at night. 37

Exposure was measured in both the traditional way using the pedestrian-vehicle volume 38

product and the exposure using PET. The ratio of the “real” exposure number over the second 39

definition was calculated to compare the exposure measures in different situations. From Table 3, 40

most of the values were less then 1 and these ratios actually varied from case to case. The 41

exposure of the number of interactions with PET less than 20 s is used in the study to compute 42

the rates and evaluate the safety performance. From the results, a higher exposures can be 43

observed in daytime compared to nighttime in most cases except for data collected at the du Fort 44

site on Wednesday, January 14th

when a hockey game brought about a large number of people at 45

nighttime, and data collected from the St-Laurent site on Thursday, January 15th

when people 46


went out clubbing. 1

From FIGURE 8 b) and d), among all interactions, a higher percentage of dangerous 2

conflicts (with PET less than 1.5 seconds) were observed at nighttime compared to daytime, 3

which indicates that pedestrians were involved in more dangerous interations with vehicles at the 4

crosswalks at night. Looking at the rates, 𝑅𝐶 values do not necessarily change from daytime to 5

nighttime, while the 𝑅𝐷𝐶 values indicate that pedestrians experience higher risks of being involed 6

in a dangerous conflict at night. 7

These results concerning the speeds and conflict rates seem to indicate that at these two 8

locations, pedestrians were at higher risks of being involved in a dangerous interaction at night, 9

when crossing speeds were on average higher. 10

However, regarding the yielding behavior of the drivers, people’s yielding compliance 11

varies from site to site. Site du Fort had a higher yielding rate at nighttime, while the yielding 12

rate is reduced at night at site St-Laurent. Upon a field inspection or Rue du Fort, in daytime, 13

vehicles were parked near the crosswalk along the sidewalk, which was free of parked vehicles at 14

night. This observation might explain the increase in yielding rate at night at this site as 15

pedestrians were easier to detect in advance by drivers. Regardless the results indicated overall 16

that the yielding compliance of the drivers at these two locations were both low (on an average 17

of 15 % - 38 % for the two sites). 18

19

20 a) Speed distribution – du Fort b) Cumulative conflict distribution – du Fort 21

22 c) Speed distribution – St-Laurent d) Cumulative conflict distribution – St-Laurent 23

24

FIGURE 8 Visualization results of the two sites – speed distributions and PET CDFs25

0%

2%

4%

6%

8%

10%

0 10 20 30 40 50 60

Per

cen

tag

e

Crossing Speed Day Night

0%

20%

40%

60%

80%

100%

0 1 2 3 4 5

Cu

mu

lati

ve

per

cen

tag

e

PET value Day Night

0%

2%

4%

6%

8%

0 10 20 30 40 50 60

Per

cen

tag

e

Crossing Speed Day Night

0%

20%

40%

60%

80%

100%

0 1 2 3 4 5

Cu

mu

lati

ve

per

cen

tag

e

PET value Day Night


TABLE 3 Results - Exposure Measures and Other Safety Measures for Daytime VS. Nighttime 1

2

3

Date Day Night Day Night Day Night Day Night

Vehicle Volume (vph) Pedestrian Volume (pph) Average Cro. Speed (km/h) Yielding Compliance

du Fort

13-Nov-14 359.4 439.2 53.8 23.3 27.95 30.57 20.83% 50.00%

14-Jan-15 -- 469.4 -- 167.8 -- 31.22 -- 38.52%

15-Jan-15 331.9 255.0 160.6 127.0 26.56 30.78 11.76% 18.18%

all 319.2 413.7 81.9 127.6 26.68 31.17 15.52% 37.66%

St-Laurent

31-Dec-14 1333.3 629.3 1087.5 423.2 24.42 32.02 32.69% 18.18%

01-Jan-15 605.0 1005.8 946.7 392.7 24.97 33.74 30.11% 25.40%

all 847.8 833.0 993.6 406.7 24.68 33.84 31.47% 22.92%

Traditional Exposure

(per hour)

Number of Interactions with

PET < 20 s (exposure)

(per hour)

No. Interactions/Trad.

Exposure (per thousand)

du Fort

13-Nov-14 19316 10247 126.9 114.2 6.6 11.1

14-Jan-15 -- 78762 -- 517.2 -- 6.6

15-Jan-15 53307 32385 213.1 208.0 4.0 6.4

all 26152 52791 158.1 306.3 6.0 5.8

St-Laurent

31-Dec-14 1450000 266323 6129.2 1758.9 4.2 6.6

01-Jan-15 572733 394988 3659.2 5280.9 6.4 13.4

all 842361 338779 4531.9 3664.3 5.4 10.8

No. Conflicts Conflict Rate No. Dangerous Conflicts Dangerous Conf. Rate

du Fort

13-Nov-14 2.5 5.0 1.97% 4.38% 0.0 0.0 0.00% 0.00%

14-Jan-15 -- 36.1 -- 6.98% -- 3.9 -- 0.75%

15-Jan-15 9.0 8.1 4.22% 3.89% 0.0 2.5 0.00% 1.20%

all 5.0 18.3 3.16% 5.97% 0.0 2.4 0.00% 0.78%

St-Laurent

31-Dec-14 55.8 27.1 0.91% 1.54% 5.0 2.9 0.08% 0.16%

01-Jan-15 37.9 44.2 1.04% 0.84% 3.3 9.1 0.09% 0.17%

all 43.9 36.4 0.97% 0.99% 3.9 6.2 0.09% 0.17%


CONCLUSION 1

This paper presented an automated video-based methodology for safety analysis for pedestrian 2

crossings at nighttime. This method was based on the use of thermal video sensors for recording 3

video in nighttime. The preliminary results showed that pedestrians were exposed to higher risk 4

levels at the study sites in nighttime as opposed to daytime conditions. The proposed automated 5

methodology can be implemented for assessing different crosswalk treatments, such as LED 6

pedestrian warning signs, an automated pedestrian detection-warning system and 7

geometric/marking treatments for improving crosswalk safety at nighttime. Results from this 8

paper showed that at the studied non-signalized pedestrian crossings, the average vehicle 9

crossing speeds are higher and percentage of dangerous conflicts were higher during nighttime 10

compared to daytime, indicating that pedestrians were at higher risks during nighttime. Not much 11

difference was found concerning the yielding compliance and the conflict rate; however, in both 12

sites the yielding compliance rate was quite low. 13

Thermal camera sensors provide a reliable solution to the limitation of common video 14

sensors in the visible spectrum when used for nighttime analysis. The main advantage of using 15

thermal cameras over regular ones is their ability to collect useful, high-quality and reliable data 16

under different environmental conditions such as in instances of low visibility and the presence 17

of glare or shadows caused by different light sources. Though the unit price of the thermal 18

camera is relatively high, rapid development of sensor technologies should bring the price down 19

and make them more accessible to institutes, research groups, governments and personal users. 20

The validation work and the potential future work about the thermal camera have been 21

discussed extensively in (20). The use of the thermal camera system for safety analysis at 22

different locations and for different types of road users in nighttime conditions will be explored. 23

The exposure used in this paper potentially provides a more precise measure to describe the 24

pedestrian-vehicle interactions which, compared to exposure measures based on traffic volumes, 25

are more closely related to pedestrian safety. A PET threshold of 20 seconds was set empirically 26

to cover all potential conflicts, while the use of this threshold needs to be further explored and 27

validated. Besides, the methodology and the safety measures used in this paper should be 28

appropriate for the analysis of signalized intersections. However, the performance of thermal 29

videos for safety analysis at busy intersections will be tested and the use of the safety measures 30

should be further validated. 31

32

ACKNOWLEDGEMENT 33

The authors would like to acknowledge the financial support provided by FQRNT and the City 34

of Montreal. In particular, we would like to thank Nancy Badeau, from the “Service des 35

infrastructures, transport et environment, Direction des Transports”. The authors recognize Taras 36

Romanyshyn for his assistance in proofreading the paper. 37


REFERENCES 1

1. NHTSA. Traffic safety facts 2013 data. National Hightway Traffic Safety Administration,

DOT HS 812 124, 2015.

2. Transport Canada. Canadian motor vehicle traffic collision statistics 2013. Canadian Council

of Motor Transport Administrators, ISBN: 1701-6223, 2015.

3. NHTSA. Traffic safety facts 2011 data. U.S. Department of Transportation, National

Highway Traffic Safety Administration, DOT HS 811 743, 2013.

4. Office of the Chief Coroner for Ontario. Pedestrian death review. Ontario, 2015.

5. Czajewski, W., P. Dabkowskib, and P. Olszewski. Innovative solutions for improving safety

at pedestrian crossings. Archives of Transport System Telematics, Vol. 6, no. 2, May 2013, pp.

16-22.

6. Hunter, W. W., J. C. Stutts, W. E. Pein, and C. L. Cox. Pedestrian and bicycle crash types of

early 1990’s. Federal Highway Administration, FHWA-RD-95-163, 1996.

7. Plainis, S., I. J. Murray, and I. G. Pallikaris. Road traffic casualties: understanding the night-

time death toll. Injury Prevention : Journal of the International Society for Child and

Adolescent Injury Prevention, Vol. 12, no. 2, 2006, pp. 125-128.

8. Rasanen, M., and H. Summala. Attention and expectation problems in bicycle-car collisions:

An in- depth study. Accident Analysis and Prevention, Vol. 30, 1998, pp. 657 – 666.

9. Huang, H., C. Zegeer, and R. Nassi. Effects of innovative pedestrian signs at unsignalized

locations - three treatments. Transportation Research Record, 2007, pp. 43-52.

10. Ryus, P., F. R. Proulx, R. J. Schneider, T. Hull, and L. F. Miranda-Moreno. Methods and

technologies for pedestrian and bicycle volume data collection. Transportation Research

Board of National Academies, Contractor’s Final Report for NCHRP Project 07-19 2014.

11. Abdel-Aty, M., and K. Haleem. Analysis of the safety characteristics of unsignalized

intersections. in 12th World Conference on Transport Research (WCTR), Lisbon, Portugal,

2010.

12. Nabavi Niaki, M., N. Saunier, L. Miranda-Moreno, L. Amador, and J.-F. Bruneau. Method

for road lighting audit and safety screening at urban intersections. Transportation Research

Record: Journal of the Transportation Research Board, no. 2458, November 2014, pp. 27-36.

13. Zahabi, S. A.H., J. Strauss, L. F. Miranda-Moreno, and K. Manaugh. Estimating potential

effect of speed limits, built environment, and other factors on severity of pedestrian and

cyclist injuries in crashes. Transportation Research Record: Journal of the Transportation

Research Board, no. 2247, 2011, pp. 81-90.

14. Tarko, A., G. Davis, N. Saunier, T. Sayed, and S. Washington. White Paper: Surrogate safety

measures of safety. in ANB20 (3) Subcommittee on Safety Data Evaluation and Analysis

Contributors, 2009.

15. St-Aubin, P., L. F. Miranda-Moreno, and N. Saunier. An automated surrogate safety analysis

at protected highway ramps using cross-sectional and before-after video data. Transportation

Research Part C: Emerging Technologies, Vol. 36, 2013, pp. 284-295.

16. St-Aubin, P., N. Saunier, L. F. Miranda-Moreno, and K. Ismail. Use of computer vision data

for detailed driver behavior analysis and trajectory interpretation at roundabouts.

Transportation Research Record: Journal of the Transportation Research Board, no. 2399,

2013, pp. 65-77.


17. Brosseau, M., S. Zangenehpour, N. Saunier, and L. F. Miranda-Moreno. The impact of

waiting time and other factors on dangerous pedestrian crossings and violations at signalized

intersections: A case study in Montreal. Transportation Research Part F: Traffic Psychology

and Behaviour, Vol. 21, 2013, pp. 159-172.

18. Zangenehpour, S., L. F. Miranda-Moreno, and N. Saunier. Impact of bicycle boxes on safety

of cyclists: a case study in Montreal. in Transportation Research Board 92nd Annual

Meeting, Washington DC, 2013.

19. Zangenehpour, S., J. Strauss, L. F. Miranda-Moreno, and N. Saunier. Are intersections With

cycle tracks safer? A control-case study based on automated surrogate safety analysis using

video data. Accident Analysis & Prevention, Vol. 86, Jan. 2016, pp. 161-172.

20. Fu, T., J. Stipancic, S. Zangenehpour, L. Miranda-Moreno, and N. Saunier. Comparison of

regular and thermal cameras for traffic data collection under varying lighting and temperature

conditions. in accepted for presentation in the 96th Annual Meeting of Transportation

Research Board, Washington D.C., 2015.

21. Shi, J., and C. Tomasi. Good features to track. In CVPR, 1994, pp. 593-600.

22. Saunier, N. Traffic Intelligence. https://bitbucket.org/Nicolas/trafficintelligence/wiki/Home.

23. Saunier, N., and T. Sayed. A feature-based tracking algorithm for vehicles in intersections. in

the Proceedings of the Computer and Robot Vision Conference, June 2006, p. 59.

24. Tang, H. Development of a multiple-camera tracking system for accurate traffic performance

measurements at intersections - Final Report. 2013.

25. Jodoin, J. P., and N. Saunier. Urban tracker: Multiple object tracking in urban mixed traffic.

in IEEE Winter Conference: Applications of Computer Vision (WACV), 2014.

26. Zangenehpour, S., L. F. Miranda-Moreno, and N. Saunier. Automated classification based on

video data at intersections with heavy pedestrian and bicycle traffic: Methodology and

application. Transportation Research Part C, Vol. 56, April 2015, pp. 161-176.

27. Peesapati, L. N., M. Hunter, M. Rodgers, and A. Guin. A profiling based approach to safety

surrogate data collection. in The 3rd International Conference on Road Safety and

Simulation, Indianapolis, USA, , 2011.

28. Gharieh, K., F. Farzan, M. Jafari, and T. Gang. Probabilistic pedestrian safety modeling in

intersections using surrogate safety measure. in ITS 21st World Congress, 2014.

29. Alhajyaseen, W. K., M. Asano, and H. Nakamura. Estimation of left-turning vehicle

maneuvers for the assessment of pedestrian safety at intersections. IATSS Research, Vol. 36,

2012, pp. 66–74.

30. Laureshyn, A. Application of automated video analysis to road user, PhD thesis. 2010,.

31. Tang, K., and H. Nakamura. Safety evaluation for intergreen intervals at signalized

intersections based on probabilistic methodology. Transportation Research Record: Journal

of Transportation Research Board (Traffic Signal Systems), Vol. 2128, 2009, pp. 226–235.

32. Qin, X., and J. Ivan. Estimating pedestrian exposure prediction model in rural areas.

Transportation Research Record: Journal of the Transportation Research Board, Vol. 1773,

2001, pp. 89-96.

33. Garder, P. Pedestrian safety at traffic signals: A study carried out with the help of a traffic

conflicts technique. Accident Analysis and Prevention, Vol. 21, no. 5, 1989, pp. 435-444.

34. Silcock, D. T., R. Walker, and T. Selby. Pedestrians at risk. in European Transport


Conference 1998, 1998, pp. 209-220.

35. Papadimitriou, E., G. Yannis, and J. Golias. Analysis of pedestrian risk exposure in relation to

crossing behavior. Transportation Research Record: Journal of the Transportation Research

Board, Vol. 10, no. 3141, 2012, pp. 79-90.

36. Hyden, C. The development of a method for traffic safety evaluation: The Swedish Traffic

Conflicts Technique. Lund Institute of Technology, Lund, ISSN 0346-6256, 1987.

37. St-Aubin, P., N. Saunier, and L. F. Miranda-Moreno. Comparison of various objectively

defined surrogate safety analysis methods. in Proc. TRB-XCIII, Washington, D.C., 2015, pp.

11-15.

2

measuring crosswalk safety at non ...n.saunier.free.fr/saunier/stock/fu16crosswalk.pdf11 midblock...

Documents